Question Answering over Knowledge Graphs via Machine Reading Comprehension

Weidong Han, Zhaowu Ouyang, Yifan Wang, Weiguo Zheng

Abstract: Due to the representation gap between unstructured natural language questions and structured knowledge graphs (KGs), it is challenging to answer questions over KGs. The existing semantic parsing-based methods struggle for building structured queries that can be executed over the KG, and thus they are difficult to cover diverse complex questions. The information retrieval-based methods suffer from poor interpretability. In this paper, we present a novel approach powered by machine reading comprehension. To transform a subgraph of the KG centered on the topic entity into text, we sketch the subgraph through a carefully designed schema tree, which facilitates the retrieval of multiple semantically-equivalent answer entities. Instead of seeking answers from all the automatically generated paragraphs, we pick out the promising paragraphs containing answers by a contrastive learning module. Finally, it is straightforward to deliver the answer entities based on the answer span that is detected by the machine reading comprehension module. The results on benchmark datasets demonstrate that our method achieves significant improvement compared with the existing methods.